In this chapter

„ Definition of VaR (parametric approach)

Chapter 5
„ Historical VaR approach
Simulation and Value at Risk --->basic concepts of VaR

„ Measuring VaR for option

1 2

Risk Measures Backgrounds

Def: Each risk measure is a single real number „ Till Guildiman, head of global research
representing different risk levels. at JP Morgan in late 1980s, creator of
Example 1: variance of the portfolio return, this the term “value at risk”.
measure tells how the variable our return is, „ The G-30 provided a venue for
but doesn’t tell the lose in monetary terms. discussing best risk management
Example 2: max possible loss, this measure is practices.
not informative, since the max possible loss is „ In July 1993, the term “value at risk”
usually the whole of the portfolio. was published in the G-30 report.

3 4
 Definition of VaR Two parameters

Value at Risk (VaR) summaries the „ The horizon period, which might be
worse loss over a target horizon with a daily, weekly, monthly, quarterly, yearly
given level of confidence. or whatever.
„ the level of confidence, which might be
90%, 95%, 99% or any other
probability we choose.

5 6

„ VaR describes the quantile of the Numerical example

projected distribution of gains and
The ABC bank 1995 daily returns
losses over the target horizon.
„ If c is the selected confidence level,
25 Mean = $5m
5% occurrence

Number of Days
VaR corresponds to the 1 - c = α lower- 20

tail level. For instance, with a 95%

15
Series1
10
confidence level, VaR should be such
5
that it exceeds 5 percent of the total 0
number of observations in the
-25
-21
-17
-13
-9
-5
-1
3
7
11
15
19
23
distribution. Daily Return

7 8
 Absolute VaR
„ Given horizon period as daily and 95% „ The former is simply the maximum
confidence, which is equivalent to amount we expect to lose with a given
regard the worst 5% daily revenues as level of confidence, measured from our
unexpected loss scenarios due to risk. current level of wealth.
„ For ABC over 1995, the value is $10
„ We can also define VaR in terms of
million, the 13th (=5%×260) worst loss
absolute dollar loss (absolute VaR), or out of 260 trading days. This $10m is
in terms of loss relative to the mean known as the absolute VaR for ABC over
revenue (relative VaR). 1995.
9 10

Relative VaR

„ The alternative VaR is measured „ Absolute VaR = -REV*

relative to the mean expected revenue „ Relative VaR
over the period. This relative VaR figure = Absolute VaR + Mean Revenue
is therefore obtained by adding the
daily mean revenues to the absolute
VaR. For ABC bank over 1995, the value where REV* is the cut-off revenues
is $10m + $5m = $15m. conditioning on the confidence level
and horizon we choose.
11 12
 „ We have
REV = R × W ⇒ REV* = R* W.
„This expression enables us to transform
„ VaR can also be represented in terms of
VaR figures in terms of revenues into VaR
the rate of return (the return) on the
portfolio. The return can be regarded as in terms of returns, which are often more
convenient to deal with.
the revenue divided by the initial value
of the portfolio. If we denote the return Abs.VaR = -R*W ⇒ Abs.VaR in return= -R*
by R, the initial portfolio value by W, Rel.VaR = -R*W + µW
and the revenue by REV. ⇒ Rel.VaR in return = -R* + µ.

13 14

Advantages of historical Advantages of historical

approach approach
„ This method is relatively simple to „ This method deals directly with the
implement if historical data have been choice of horizon for measuring VaR.
collected in-house for daily marking-to- For instance, to obtain a monthly VaR,
market. The same data can be stored the user could re-construct historical
for later reuse in estimating VaR. monthly returns over, say, 5 years.
„ Perhaps most important, it does not
depend on the probability distribution
we assume for the return.
15 16
 Problems of historical „ Historical approach will be very slow to
incorporate structural breaks, which are handled
approach more easily with an analytical methods such as
RiskMetrics.
„ The historical approach requires a “sufficient
„ It may not be suitable for analyzing derivatives
history” of price changes. To obtain 1000 risk. Derivatives products can have a very short
independent returns of a 1-day move, we life, say 3 months, in the market. Therefore, one
require 4 years of continuous data. If monthly may want to make use of the underlying asset
return is concerned, 1000 independent prices. However, incorporating the prices of
returns stand for more than 80 years data. underlying assets can be terribly difficult because
„ Only one sample path is used so that it no model is assumed to relate underlying asset
generates difficulties for scenarios analysis. price and the derivatives price. Otherwise, the
distribution-free advantage of the method is
broken.
17 18

„ The method puts the same weight on all „ The method becomes cumbersome for large
observations, including old data points. portfolios with complicated structures. In
„ It could produce a very large error in estimating practice, users adopt simplifications such as
VaR. For instance, a 99% daily VaR estimated grouping interest rate payoffs into bands,
over 100 days produces only one observation in which considerably increase the speed of
the tail, which necessarily leads to an imprecise computation. But if too many simplifications
VaR measure. Thus very long sample paths are are carried out,the benefit of distribution-free
required to obtain meaningful figure. The valuation can be lost.
dilemma is that this may involve observations
that are not relevant.
19 20
 Parametric VaR: Normal
Returns
assumption
„ The parametric VaR focuses on the „ Arithmetic return
random process that describes the RtA = (Pt+1 – Pt)/Pt = δPt /Pt ≥ -100%
behavior of the asset's (portfolio's) return. where Pt is the price of the asset at time t.
That is we make assumption about the „ This implies that the return does not follow
probability density function (pdf) of return. a normal distribution since a normal random
„ One common approach considers the variable can take any real number.
normal distribution for the return process.
„ Does the return really follow normal ?
21 22

Returns

„ Geometric return „ Ito’s lemma verifies that

RtG = ln[Pt+1/Pt]=lnPt+1- lnPt = δlnPt mean of RG ≅ mean of RA - ½(variance)
„ It can be any real number. (good!) and
„ The differential forms for arithmetic variance of RG ≅ variance of RA
return and geometric return are the same.
dlnPt/dPt = 1/Pt ⇒δlnPt ≅ δPt /Pt ⇒RtG ≅ RtA.

23 24
 Ito lemma Summary

„ Recall: „ We know that the geometric return can be

assumed to follow a normal distribution.
Suppose dP/P = µdt + σdZ, dZ ∼N (0,dt).
„ If the mean and variance for the geometric
return are found to be µG and σG2, respectively.
„ Applying Ito lemma, we have Then, we should aware that
dlnP = (µ - σ2/2) dt + σdZ. σA = σG
µA = µG + σG2 /2

25 26

Normal VaR in return Normal VaR in return

Suppose we want to determine the absolute Consider R ~ N (µ, σ2). We can always
VaR = -R* with confidence level c (or tail transform R into a standard normal
probability α = 1 - c). Given the pdf for random variable:
the return, we should be able to compute
the cut-off return by R−µ
Z = ~ N ( 0 ,1)
Pr[R < R*] = α. (6.8) σ

27 28
Hence, (6.8) becomes „ For instance, zα = -1.65 for α = 5% (95%
confidence) and

⎛ R * −µ ⎞ ⎛ R * −µ ⎞
Pr ⎜ Z < ⎟ = N⎜ ⎟ =α
⎝ σ ⎠ ⎝ σ ⎠ Area = 5%

R * −µ
= N −1 (α ) = zα
σ
-1.65 0 x

29 30

„ zα = -2.33 for α = 1% (99% confidence)

„ This gives us the normal VaR measures:

Absolute VaR in return
Area = 1%
= -R* = -σ zα - µ
Relative VaR in return
= -R* + µ = -σ zα
x
-2.33 0

31 32
 „ It is possible that the geometric return
„ As the variances computed from is used for computation because of the
arithmetic returns and geometric returns normality assumption.
are the same, two approaches produce „ But, the user think that the arithmetic
the same relative VaR. return or the VaR in revenue would be
„ One should aware of the absolute VaR much more meaningful in practice.
since the mean return appears in the Then basically, we have two ways to
expression. deal with it.

33 34

1. If you would like to see the VaR in 2. For VaR in arithmetic return, consider
revenue only, you can apply the the relationship of (6.7). We have
definition for the geometric return to Absolute VaR in arithmetic return
retrieve the VaR in revenue and = -σ zα - (µG + σ2/2)
(RG)*=ln(Pt*/W) ⇒ Pt* = W exp[(RG)*] Abs. VaR in revenue
⇒ REV* = Pt* - W. = Abs. VaR in arithmetic return × W

35 36
 „ Provided that daily returns are
Holding period adjustment independently distributed from one day
to the next, the mean return over the
„ If we are given information based on 20-day period is:
one particular holding period, but want µmonthly = 20µdaily
VaR information based on a different „ Similarly, the variance of the 20-day
holding period. return is:
„ Imagine we are still dealing with daily σ2monthly = 20σ2daily
return, but are now interested in a ⇒ σmonthly = σdaily√20
longer horizon period of, say, 20 days
(i.e. one business month).
37 38

„ The above formula indicates that the market

„ The VaR for the longer holding period is risk of a position is increase with volatility,
then : initial investment cost, the square root of the
VaRmonthly = -zα σdailyW √20 holding period and the users' risk averseness.
= √20 VaRdaily „ Given the same portfolio, a more risk averse
„ In general, people present the normal investor would like to have a larger
(relative) VaR measure as confidence in the portfolio's performance.
VaRTα (relative) = -zα σyearly √T × W „ Hence, a smaller α is chosen. A smaller α
gives a greater VaR. Do you think it matches
with your understanding of market risk?

39 40
 Does return follow normal? Checking the distribution
„ Checking the distribution
„ Eyeball testing: Q-Q plot „ Kolmogorov-Smirnov Goodness of Fit:
Let Fn(x) to be the empirical distribution
and F0(x) to be the hypothesized dist. The
Kolmogorov-Smirnov statistic:
Dn = supx|Fn(x) – F0(x)|.
Test: H0: F(x) = F0(x) vs H1: F(x) ≠ F0(x)

Statistical Software (S-Plus) provides details for

the above test.
√ X
41 42

How to calculate VaR from a

An example proposed distribution?
Data vs normal Data vs t(2)

„ Ifwe have strong statistical

knowledge, we may obtain the VaR
explicitly through hand-calculation.
„ Usually, hand-calculation fails.
The p-values of KS test:
„ Simulation is the unique approach.
0.7942 0.9610

43 44
Monte Carlo Simulation to VaR Understanding MCS
MC method estimates VaR on the basis of
simulation, that means to produce possible Example: Suppose the geometric return of a
future events, derived from statistical stock follows normal:
models. It involves a number of steps. ln(St+1)-ln(St) = RGt = µt + σtεt,
„ Step 1: Select an appropriate statistical where εt ~ N(0,1). By drawing 1000 εt’s, we are
model for the asset(s). able to obtain two information:
„ Step 2: Construct fictitious price paths for the (1) 1000 possible stock prices in the future.
random variables involved. (2) The 51st smallest return is the 95% VaR
„ Step 3: Repeat Step 2 for enough times. after sorting the data in ascending order.
„ Step 4: Read off the VaR from the proxy
distribution. 45 46

Some extensions Advantages of Simulations

The above idea can easily be extended to „ The implementation is simple.
handle the following situations: „ They are widely used in financial
„ εt follows a t-distribution, which may be a sectors.
better one to fit the data. „ They can easily handle the price risks
„ A portfolio contains two assets. However, the associated with non-linear positions. Eg.
covariance among assets should be
considered during the simulation procedure.
Exotic derivatives position.
„ The prices of derivatives and fixed income
„ They can provide estimation error of
securities, like options and bonds, can also be VaR for any distributions.
obtained. „ They can generate future scenarios as
47 much as you want. 48
 A heavy tail distribution The shape of GED

„ Generalized Error Distribution (GED):

Rt = µ t + σ t ε t
where ε follows a GED with density function
given by:
f (ε t ) =
(
ν exp − 12 ε t / λ
ν
)
λ 2(1+ν ) Γ(ν −1 )
−1

and 1/ 2
⎡ 2 − ( 2 /ν ) Γ(1 / ν ) ⎤
λ=⎢ ⎥
⎣ Γ(3 / ν ) ⎦
It becomes normal when v = 2.
49 50

A case study Normality test

„ I download 10-year daily closing prices of
Dow Jones Industrial Index (DJI) and convert
them into 2,624 daily returns.
„ Sample mean = 0.04%, Sample volatility =
1.16%.
„ 95% normal VaR = 1.87%
„ 99% normal VaR = 2.66%
„ Using the sample mean and sample volatility,
I convert the data into standardized returns.
„ Then, I construct the qq-plot (normality plot).

51 52
 Improvement Estimation

„ I think that normal may not be a good „ The log-likelihood function is

N
assumption for the data so that I consider the l (ν ) = ∑ log f ( zi | ν ).
GED. i =1
„ The GED requires an additional parameter ν „ Then, I search for a suitable ν such that the
log-likelihood function is maximized.
to estimate. I employ the maximum likelihood
„ My computation shows that the MLE for ν is
estimation. 1.21, meaning that the empirical distribution
„ Denote f(z|ν) be the GED density given in has a heavy tail.
page 49, where z is the standardized return „ To verify if there is an improvement, I
(observations). construct the qq-plot again.

53 54

Testing the data GED-VaR

„ The data shows that GED(1.21) is more
appropriate to describe the DJI returns.
„ In other words, it may be a better idea to
obtain VaRs from the GED(1.21).
„ The problem is we do not have a closed form
solution for the GED quantiles.
„ The unique way to compute VaR relies on MC
simulation.
„ The technical issue is how GED random
variables are generated.

55 56
 Idea of generating GED Proposal distribution

„ Inverse transform is impossible since „ The proposal distribution should satisfy

we have to deal with an improper two conditions:
integral. „ It shares the same domain as the target
„ We may have to use the rejection distribution. In this case, the target
distribution is GED.
method.
„ Its tail must be heavier than that of the
„ Question: What is the proposal
target distribution, implying that the
distribution? normal density cannot be employed.

57 58

Proposal distribution 2f(x)/g(x) vs. x

„ Since it is GED(1.21), the tails are heavier than
normal but thinner than double exponential,
exponential distribution is a possible proposal
distribution for the positive GED.
„ I used Exp(1) distribution to generate positive GED.
„ To implement rejection method, we need the
constant c which bounds the function f(y)/g(y) for all
y, where f is the target and g is the proposal.
„ Obtaining the constant c with differentiation is
sometimes cumbersome. I use a graphical approach
in practice.

59 60
 Rejection method Results

„ From the graph, I choose c = 1.2. „ Using the rejection method, I generate
„ The rejection method runs as follows: 1 million possible returns.
„ Generate Y ~ Exp(1), i.e. Y = -ln(U1) „ From the generated sample, I obtain
„ Generate U ~ U(0,1) that 95% VaR = 1.87 (same as the
„ If U < 2f(Y)eY/1.2, then set X = Y. normal VaR) and that 99% VaR =
Otherwise, go to step 1. 3.04% > 2.66% (the normal VaR).
„ Generate V ~ U(0,1) „ This example shows that normal VaR is
„ If V < 0.5, then X = -X. only applicable for less confidence level.

61 62